TL;DR
This paper introduces a reproducible framework for generating heterogeneous federated datasets from existing centralized survival analysis data. By using Dirichlet-based splitting algorithms, researchers can now precisely control data imbalance and distribution shifts among simulated clients, providing a much-needed benchmarking tool for developing privacy-preserving survival models.
What Problem Does This Paper Address?
Survival analysis—predicting the time until an event like a mechanical failure or a medical complication—is critical in fields where data is often siloed and highly confidential. While Federated Learning (FL) offers a way to train models on this distributed data without compromising privacy, the field lacks standardized, realistic datasets for testing.
In the real world, data isn't "Identically and Independently Distributed" (IID). One hospital might have thousands of patients (quantity skew), while another might specialize in high-risk cases with much shorter survival times (label skew). Most existing FL simulations use uniform data splits that don't reflect these messy realities. Without a way to simulate this "heterogeneity," we can't know if our federated survival models will actually work when deployed in diverse, real-world environments.
Methods
The authors propose two novel splitting algorithms based on the Dirichlet distribution, which acts as a mathematical "knob" to control how skewed the data becomes.
- Quantity-Skewed Splitting: This assigns a different number of samples to each client. By tuning a similarity parameter ($\alpha$), researchers can simulate a federation where some clients are "data-rich" and others are "data-poor."
- Label-Skewed Splitting: This is more complex. The survival timeline is divided into discrete "bins." The algorithm then ensures that different clients receive different proportions of samples from these bins. For example, one client might end up with mostly "early-event" samples, while another gets "late-event" or "censored" samples.
The beauty of this approach is that a single hyperparameter ($\alpha$) allows researchers to transition smoothly from a perfectly uniform distribution ($\alpha \to \infty$) to extreme heterogeneity ($\alpha \to 0$).
Datasets
The study applies these splitting techniques to five well-known centralized survival datasets, effectively turning them into federated benchmarks:
- GBSG2: Breast cancer study data.
- METABRIC: Molecular taxonomy of breast cancer.
- AIDS: Clinical trial data for HIV treatment.
- FLCHAIN: Assay of serum free light chain.
- SUPPORT: A large study on hospitalized adults' prognoses.
These datasets vary significantly in size (from 686 to over 9,000 samples) and censoring rates (28% to 68%), providing a diverse testing ground.
Key Insights and Ideas
- Quantity Skew $\neq$ Distribution Skew: One of the most important findings is that simply giving clients different amounts of data (quantity skew) does not significantly change the underlying survival curves. The "heterogeneity score" remained below 5% in these cases. This means quantity skew tests a model's resilience to imbalance, but not to population differences.
- The Power of Time-Binning: By treating survival times as "classes" through binning, the authors successfully adapted Dirichlet-based partitioning—usually used for image classification—to the continuous time-domain of survival analysis.
- Quantifying "Difference": The authors didn't just look at the splits; they used log-rank tests to statistically prove that their label-skewing algorithm creates populations that are truly different from one another.
Key Results
The authors introduced a heterogeneity score ($h$), representing the percentage of client pairs that have statistically significant differences in their survival distributions.
- Label-Skewed Success: For the
SUPPORTdataset with 5 clients, the $h$ score jumped from 0.5% (at $\alpha=1000$, near-uniform) to 94.0% (at $\alpha=0.1$, highly heterogeneous). - Visual Evidence: Kaplan-Meier plots (which visualize survival probability over time) clearly showed the curves for different clients diverging as $\alpha$ decreased, mimicking real-world silos where patient outcomes vary by location or demographic.
- Reproducibility: The implementation is publicly available, allowing other researchers to generate the exact same splits for their own model comparisons.
Limitations and Future Work
While this paper provides a robust framework for data heterogeneity, it focuses primarily on right-censoring (where we know an event hasn't happened yet, but not when it will). Future work could extend these algorithms to handle left-censoring or interval-censoring.
Additionally, the paper focuses on data heterogeneity. In real-world FL, system heterogeneity (differences in client hardware, battery life, and network speed) is also a major factor. Combining this data-splitting framework with system-level simulation would be the "gold standard" for future federated survival analysis research.
Summary updated